The AI “adviser” algorithm monitors the performance of other machine learning algorithms as autonomous experiments progress. It provides human scientists with insights that inform refinements to experimental plans.
The rapid global growth of electronics poses significant challenges for materials innovation. Traditionally, discovering new electronic materials can take many years. Emerging technologies, such as flexible electronics and bioelectronics, require much faster development. Artificial intelligence (AI)-driven autonomous experimentation has the potential to significantly speed the discovery of breakthrough materials.
However, AI algorithms need to be trained with large amounts of data to make good decisions. A lack of data on electronic materials has limited the effectiveness of AI-driven experiments. This data scarcity is due to the long time it takes to design, fabricate and evaluate materials. New strategies are needed that enable AI systems to perform effectively with small datasets.
A research team led by the U.S. Department of Energy’s (DOE) Argonne National Laboratory developed an innovative AI “adviser” that addresses these challenges. The AI adviser is a specialized algorithm that monitors the performance of machine learning algorithms as autonomous experiments progress. It also extracts key insights. Human scientists use these insights to refine their hypotheses or experimental plans.
This AI-human collaboration boosts efficiency and increases the likelihood of discovering high-performance materials. The AI adviser is inspired by robo-advisers that manage financial portfolios.
The researchers applied the adviser to an AI-driven investigation of electronic materials called mixed ion-electron conducting polymers (MIECPs). The study yielded important findings on how the packing structure of these materials influences their performance.
“AI algorithms used in autonomous laboratories lack the ability to make adaptive changes to experiments based on small datasets,” said Jie Xu, one of the study’s lead authors. Xu is an Argonne scientist and assistant professor of molecular engineering at the University of Chicago Pritzker School of Molecular Engineering. “The AI adviser transformed our robotic laboratory from a relatively static workflow into a highly adaptable one. The results were compelling.”
Xu added, “I expect researchers to apply our adviser concept and methods to various materials. This will help accelerate new discoveries.”
The study was published in Nature Chemical Engineering. In addition to Argonne, the research team included the University of Chicago, DOE’s Lawrence Berkeley National Laboratory (LBNL), the University of Southern Mississippi and the University of Central Florida.
Human-AI Collaboration Drives Successful Study
The scientists integrated the AI adviser into Polybot. This AI-guided robotic laboratory is in the Center for Nanoscale Materials, a DOE Office of Science user facility at Argonne. Polybot’s robotic platforms autonomously synthesize and characterize materials. Then, its AI algorithms analyze the experimental data. Based on this analysis, the algorithms decide on the next experiments.
To test the adviser, the team used Polybot to investigate MIECPs. These soft organic materials can conduct both electrons and ions simultaneously. This dual conductivity makes them promising for a wide range of applications, including wearable electronics and energy storage.
The researchers wanted to uncover the design principles that govern performance of MIECPs and to optimize their properties. Their approach was to evaluate the materials in fabricated transistors.
Polybot autonomously deposited the MIECP onto substrates. Then, it fabricated transistors from the material, measured the transistors’ performance and characterized the material’s properties. The key performance metric quantified the transistors’ ability to move and store electronic and ionic charge.
In experimental iterations, Polybot’s AI algorithms varied the processing conditions. These conditions include the material’s concentration in solution, deposition temperature, deposition speed and substrate features. Two goals informed these adjustments: The first was to explore diverse processing conditions with as few experiments as possible. The second was to maximize understanding of material structure-property relationships.
An Algorithm Supervisor
The adviser is an AI algorithm that “supervises” Polybot’s AI algorithms. It evaluated experimental progress, analyzed experimental datasets and compared the algorithms’ performance. It communicated key patterns and trends to human scientists via a livestreaming platform. The idea was to inform the scientists’ decisions on refinements to experimental workflows and parameters.
At one point, the adviser observed diminishing performance improvements. It suggested switching to another AI algorithm for subsequent experiments. The scientists implemented the recommendation, and device performance improved significantly.
The adviser also found that deposition speed contributed the most to improved performance. It shared this insight with the scientists. This informed their decision to widen the scope of investigation for this parameter, driving additional performance gains.
The adviser-enabled experimental adaptations allowed Polybot to complete the study with just 64 experiments. This is remarkably fast considering that there were more than 4,300 possible combinations of processing conditions. The study yielded a diverse dataset on material structure-property relationships.
In-Depth Characterizations Yield Valuable Design Principles
The research team also used lasers, X-rays and electric current to characterize the structure and properties of the 10 most representative material samples. The objective was to better understand how the samples’ packing structures influenced their performance. Some of the characterization work was performed at the Advanced Light Source, another DOE Office of Science user facility at LBNL.
These in-depth characterizations revealed two structural features that played a crucial role in better performance: wider spaces between layers and thinner fibers. The team also discovered that the material crystallizes into two distinct structures. These significant findings can be leveraged to design higher-performing MIECPs.
The study was supported by the DOE Office of Science, Basic Energy Sciences program.